6-Week AI System Design Challenge: Practice Backend Design with AI Agents + Prepare for Technical Interviews

AI writing code for you is a given now. The problem is that deciding whether that code fits our situation is still a human responsibility—and CLAUDE.md is where you write down that judgment. Every week, you give an AI coding agent a real-world backend problem. However, you are not given a completed set of standards. You start with a thin CLAUDE.md containing only the requirements, find the places where the agent violated the standards using a grading rubric, turn those places into prohibitions with reasons, add them back in, and run it again. v1 (requirements) → 1st run → grade deviations → add constraints (v2) → 2nd run → verify → critique The materials come from real-world work: zero overselling for first-come, first-served coupons, exactly one payment even when retries run wild, not being off by even 1 won when settling 100,000 events per second, and designing blue-green deployments and alerts. The passing criteria are not my opinions—they are test results and k6 numbers. What you submit is not the code, but the CLAUDE.md you developed and a record of “what was caught.” ── How this is used in interviews Each week’s assignment is paired with a real interview question. When asked, "How would you design a system for 100,000 people trying to claim a first-come, first-served coupon at the same time?", someone who has completed this challenge can answer like this: "I moved the inventory to Redis and combined the check and decrement into a single atomic operation. I hammered it with 200 VUs for 20 seconds and confirmed a p95 of 14 ms and zero overselling. JVM locks become meaningless the moment you have two servers, so I didn’t use them." These aren’t memorized answers—they’re numbers you measured yourself. This can also be used in hiring processes that allow AI. What those processes look for is not "Did you use AI?" but "How did you instruct it and verify the results?" The "My CLAUDE.md" you write in Week 6 becomes a portfolio piece as-is. It is not a clone of someone else’s project, but a document containing your own judgment about your own code, making it impossible to replace. ── For people who · Are 1–3-year backend developers who can build features with AI but get stuck when code review asks, "Why did you do it this way?" · Are responsible for concurrency, payments, or deployments but feel anxious because they have never experienced an outage · Have no idea what or how to instruct an agent ── Prerequisites JDK 21 · Docker Desktop · k6 · Any AI coding agent (the free version is fine) At least one experience with Kotlin/Java syntax and Spring Boot. No prior Redis or k8s experience is required. Local Kubernetes (Docker Desktop’s Kubernetes or kind) is required only for Week 6. ── Explanations The seven AI Hands-on lessons from the course “I’ll Fill in the Experience for You: Practical System Design for Juniors in the AI Era” serve as explanations for the weekly missions. Try them yourself first, then watch the videos afterward. ※ Algorithm coding tests are not covered. This is a challenge about design judgment and verification.